Robustness checks in xtabond2

[Marco Astori]

Dear statalists, I'm using xtabond2 and I'd like to know why my lagged variable, when I use the collapse option becomes insignificant, whereas without it it doesn't.
I read Roodman 2009 and I think I've understand the option: it generates instruments for each lag, but not for eeach lag and period, so there are 1+1+1+1... moment conditions instead of 1+2+3+4+...T-1 moment conditions. Is that correct?

[Sebastian Kripfganz]

Have a look at slides 12 and 19 of my 2019 London Stata Conference presentation for a comparison of instruments with/without collapsing:

• Kripfganz, S. (2019). Generalized method of moments estimation of linear dynamic panel data models. Proceedings of the 2019 London Stata Conference.

[Marco Astori]

Have a look at slides 12 and 19 of my 2019 London Stata Conference presentation for a comparison of instruments with/without collapsing:

• Kripfganz, S. (2019). Generalized method of moments estimation of linear dynamic panel data models. Proceedings of the 2019 London Stata Conference.

Thank you Sebastian Kripfganz if I've understand correctly with the collapse option, we verify the summation of moment conditions and not singular moment conditions that generate that sum.

Is it possibile that with the collapse option my lagged dependent variable becomes not significant in a correct model?

[Sebastian Kripfganz]

Yes, your understanding is correct. Theoretically, the summation could still yield a valid moment condition even though some of the individual moment conditions are invalid. In practice, however, such a peculiar case is hard to imagine. Having said that, it could still happen that the Hansen test rejects e.g. the validity of the collapsed instruments while not rejecting the individual instruments. This discrepancy could most of the time be attributed to finite-sample performance issues of the test, e.g. in relation to a too-many-instruments problem.

Similarly, the estimated coefficients and standard errors can differ (and sometimes unfortunately to a substantial degree), even though the underlying model did not change substantially. So, yes, what you are observing is certainly possible but it is hard to give a more enlightening explanation.

[Marco Astori]

Thank you very much Sebastian Kripfganz